Constraints and resources · Stamped architecture

Purpose#

“Code withholds on a known constraint conflict” is only true if code can evaluate the constraint. Plant constraints are mostly unwritten. This doc defines a minimal typed vocabulary, how rows are stored, how the evaluator works, and how tacit constraints become typed without an LLM deciding the gate.

Constraint kinds (registry)#

Admitted when a family or pattern uses one:

Kind	Intent
`forbid`	Named action / state must not occur in scope
`must-run`	Asset or resource must remain available / running
`time_window`	Only valid inside / outside a window
`bound`	e.g. feeder max kW
`cumulative_capacity`	Shared resource capacity over a window
`mutual_exclusion`	Two footprints cannot both hold
`precedence` / min separation	A before B, or minimum gap
`reserve_margin`	Headroom that must remain
`availability`	Crew, material, fixture
`do_not_disturb_order`	Order / due as do-not-disturb context

New kinds are registry entries + evaluator binding + replay — not a kernel edit (21-registries-and-stage-graph.md).

Constraint row#

Field	Meaning
scope	asset, flow edge, shared resource, area, plant
applicability	mode, product, shift, state
effective_time / recorded_time	bitemporal
severity	`hard` · `conditional-hard` · `advisory`
owner	named plant role
source	site pack, owner edit, proposed-from-rejection
expiry	optional
version	row version

Evaluator (code)#

Returns satisfied | violated | unknown, plus the smallest conflicting fact set, recorded in the DecisionTrace.

Result	Kernel
`violated`	`withhold`
`unknown` on hard / conditional-hard treated as hard for the neighbourhood	`withhold` (`reason=constraint_unknown`)
`satisfied`	continue

The Constraint analysis model explains results and may flag possible conflicts. It never decides satisfied / violated / unknown. An LLM “possible conflict” without a code result → withhold (00-kernel.md).

Missing structure needed for a cross-asset check → unknown (02-plant-structure.md).

Tacit constraints#

Rejection reasons and owner edits become proposed constraint rows, routed to the named plant owner. Most real plant constraints are unwritten; this is how they become typed. Proposed rows are advisory until confirmed. Rejected proposals are remembered (same spirit as topology suggestions).

Action footprint#

Every candidate and open card carries:

assets
shared resources
crew / role
material
time window with start, end, and lag

Portfolio conflict = footprint overlap and/or constraint evaluation conflict (09-portfolio.md). The claims index in the PSM holds open-card footprints for that check.

Hard vs soft (do not confuse)#

	Constraint severity	Kernel gate
Hard constraint violated / unknown	severity on the row	Hard gate — never tunable
Soft gates (agreement, attention, …)	not constraint kinds	Soft gate registry (`22-missed-opportunities.md`)

Advisory constraint severity may inform ranking and explanations; it does not by itself force withhold unless a family maps it into a hard neighbourhood.

Why typed#

Constraint-programming practice (precedence, no-overlap, cumulative resources) covers most shared-resource cases without a scheduler or a plant ontology. Prose in a prompt is not a gate.

v1 slice vs later#

v1	Later
Kinds needed by Pilot families (forbid, bound, time window, availability, do-not-disturb, mutual exclusion where topology exists)	Full kind set as families commission
Code evaluator + unknown→withhold on hard	Same
Tacit → proposed rows → owner confirm	Same; richer suggestion UX
Advisory severity informs rank/explain only	Optional family maps of advisory→hard neighbourhoods